开源大语言模型Qwen 3.8 27B在单台工作站上成功完成了商业应用许可证校验系统的逆向工程任务，耗时约30分钟1。该模型通过静态分析恢复了隐藏的加密密钥、理解了整个认证架构，并生成了有效的认证绕过方案1。这一案例表明27B规模的本地开源模型已能执行曾属于前沿模型的复杂任务1。
该测试在Lenovo ThinkStation PGX工作站上进行，配置为128GB统一内存、273GB/s带宽1。模型运行时占用17GB VRAM1，基础推理速度为15-30 tokens/s，经优化后代码推理速度可达50 tokens/s1。模型最初拒绝了越狱提示，但在理解了安全的分析边界后改变了态度并完成任务1。在分析过程中，模型自动发现并纠正了第一次密钥恢复中的哈希不匹配错误1。整个分析过程完全离线进行，未涉及任何云服务1。
A user running the open-source Qwen 3.8 27B model on a local workstation tasked the AI with reverse-engineering a commercial application's license validation system, and the model completed the assignment in approximately 30 minutes 1. Through static analysis, the model successfully recovered hidden encryption keys, understood the entire authentication architecture, and generated a functional authentication bypass 1. The demonstration underscores that 27-billion-parameter open-source models have become capable of executing complex tasks previously confined to cutting-edge proprietary systems 1.
The reverse-engineering work was performed entirely offline on a Lenovo ThinkStation PGX workstation equipped with 128GB unified memory and 273GB/s bandwidth, with the model consuming 17GB of VRAM 1. The model's inference speed ranged from 15–30 tokens per second at baseline performance, accelerating to 50 tokens per second when optimized for code reasoning tasks 1. During the analysis, the model initially rejected jailbreak prompts but shifted its stance after understanding the legitimate boundaries of security analysis and proceeded to complete the assignment 1. The model also autonomously identified and corrected an error during the first key recovery attempt, discovering a hash mismatch and resolving it without external intervention 1.
评论
还没有评论，欢迎留下第一条。